How to Make ChatGPT More Accurate (Settings + Prompts)

ChatGPT doesn’t have a single “be accurate” switch. What it has is a handful of settings that each close one gap: guessing, stale information, sloppy reasoning, or answering from memory when it should be reading your document. Set them once and add a few prompts, and you’ll get noticeably fewer confident wrong answers.

Here’s each setting, where OpenAI’s help center says to find it, what it fixes, and the prompts to use on top. OpenAI moves menus around often, so if a label looks different on your screen, look for the closest match.

The quick setup: seven settings and what they fix

1. Add an accuracy rule to your custom instructions

Custom instructions are standing guidance ChatGPT considers in every chat. On the web, OpenAI’s custom instructions help page says to open Settings, select Personalization, make sure “Enable customization” is on, and type into the Custom Instructions field. In the mobile apps the same controls sit in your settings, sometimes labeled Customize ChatGPT. They’re available on every plan. As of September 2026, Free and Go users get up to 1,500 characters and paid plans up to 5,000.

This block fits comfortably inside the free limit:

Accuracy matters more to me than speed or sounding confident.
- If you don't know something or can't verify it, say so plainly. "I'm not sure" is a good answer.
- Never invent sources, links, quotes, statistics or product features. Only cite things you can link.
- For anything that changes over time (prices, laws, software, people's roles), search the web or tell me your information may be out of date.
- Separate facts from your inferences and opinions.
- If my question assumes something false, tell me before answering.
- If I push back, change your answer only if I give new evidence or a better argument.
- If the answer depends on details you don't have, ask me one clarifying question first.

Accuracy matters more to me than speed or sounding confident.
- If you don't know something or can't verify it, say so plainly. "I'm not sure" is a good answer.
- Never invent sources, links, quotes, statistics or product features. Only cite things you can link.
- For anything that changes over time (prices, laws, software, people's roles), search the web or tell me your information may be out of date.
- Separate facts from your inferences and opinions.
- If my question assumes something false, tell me before answering.
- If I push back, change your answer only if I give new evidence or a better argument.
- If the answer depends on details you don't have, ask me one clarifying question first.

While you’re on that page, you’ll see Personality and Characteristics settings too. They’re worth setting to taste, but they change tone, not truthfulness. OpenAI describes Characteristics as a way to “keep the same capabilities, but adjust the vibe.” For more on why the “don’t cave” line matters, see making your AI more accurate and honest.

2. Clean up what ChatGPT remembers about you

Memory helps ChatGPT skip the repetition, but stale memories quietly skew answers. OpenAI’s Memory help page notes that memory can even shape how ChatGPT words its web searches, so an old city or job title can pull in the wrong local results.

• Review it: Settings, then Personalization, then Memory. Depending on your account, you’ll see controls like a memory summary, saved memories, and Reference chat history.
• Ask directly: type “What do you remember about me?” and correct anything wrong or outdated. You can tell ChatGPT a memory is incorrect, or delete it.
• Get a neutral answer: start a Temporary chat and choose Unpersonalized before your first message. It won’t use your memories or custom instructions.

Memory should hold facts about you that stay true. If you want to set up a proper profile, this nine-line setup shows what to include.

3. Make it search for anything current

ChatGPT’s built-in knowledge stops at a training cutoff, so anything that changes needs a live search. OpenAI says ChatGPT may search automatically when your question would benefit from current information. To force it, open the tools menu, choose Search, or type / in the message box and pick Search. If an answer came back without searching, you can regenerate it with “Search the web” where that option appears. Web search is available on every plan, including Free.

Then do the part the feature can’t do for you. OpenAI’s own help page says search results and citations “can be incomplete, outdated, or incorrect” and tells you to open the cited source to check it. For a fast way to do that, use the 3-minute AI fact-check workflow.

4. Pick a thinking level for hard questions

For multi-step problems (comparisons, math, research questions, anything with a chain of logic), a reasoning mode helps. OpenAI has said its reasoning models hallucinate less, though not never.

As of September 2026, the model picker offers Instant for quick everyday answers. Paid plans get a Thinking slider with higher levels, and Free and Go users get a Think option. One detail catches people out: according to OpenAI’s model help page, on Plus and Pro, Instant doesn’t automatically switch to a higher thinking level, and asking it to “think harder” won’t change the level. You have to choose it in the picker. Labels change often, so check the picker itself.

Use Instant for rewrites and quick lookups you’ll verify anyway. Switch up for anything where one wrong step ruins the answer.

5. Upload the source instead of asking from memory

The single biggest accuracy gain is giving ChatGPT the document and asking about that text. File uploads work on Free and paid plans, within usage limits. Ask it to quote the exact lines it relies on, so you can find them yourself.

One gotcha from OpenAI’s file FAQ: outside ChatGPT Enterprise, document files are read as text, and ChatGPT discards images inside them. A chart or scanned table in a PDF may be invisible to it. If the key numbers live in a chart, paste them in as text or upload a screenshot of the chart as an image.

6. Put recurring work in a Project

Projects keep related chats, files and instructions together, so ChatGPT works from the same sources every time instead of whatever you remembered to paste. Create one with New project in the sidebar, add your reference files, and add instructions from the project’s settings.

Here’s the catch: OpenAI says project instructions apply only within that project and override your global custom instructions. If you rely on the accuracy rule from step 1, paste it into each project’s instructions too.

7. Use deep research when one search isn’t enough

For questions that need many sources compared, start Deep research from the tools menu (+). It proposes a research plan you can edit, runs for a while, and returns a structured report with citations. You can also restrict it to specific websites under Sites, then Manage sites, which is useful when you only trust official sources. Usage limits depend on your plan, and the citations still need clicking.

Prompts that make ChatGPT more accurate

Settings set the defaults. These prompts handle the moments that matter.

For questions about a document:

Answer using only the attached file. Do not use outside knowledge.
For each point, quote the exact sentence from the file that supports it.
If the file doesn't answer part of my question, say "not in the file" instead of filling the gap.

Question: [your question]

Answer using only the attached file. Do not use outside knowledge.
For each point, quote the exact sentence from the file that supports it.
If the file doesn't answer part of my question, say "not in the file" instead of filling the gap.

Question: [your question]

For questions about anything current:

Search the web and answer [your question].
- Prefer official sources: the company, the government agency, or the original publisher.
- Link a specific page for every factual claim and note its publish or update date.
- Say "I couldn't verify this" rather than guessing.
- End with the 2 claims you're least sure about.

Search the web and answer [your question].
- Prefer official sources: the company, the government agency, or the original publisher.
- Link a specific page for every factual claim and note its publish or update date.
- Say "I couldn't verify this" rather than guessing.
- End with the 2 claims you're least sure about.

Before you use any answer that matters: run it through the free AI hallucination checker prompt, which lists every claim and tells you which ones to verify.

Mistakes that keep ChatGPT inaccurate

• Assuming a paid plan fixes it. Paid plans add higher thinking levels and bigger limits. Neither makes ChatGPT right about things it never learned. (Weighing an upgrade? See ChatGPT Free vs Plus.)
• Letting stale memories steer answers. Review Memory every few months.
• Typing “think harder” and expecting a different mode. Pick the level in the model picker.
• Trusting a citation without opening it. A link is a claim, not proof.
• Forgetting that project instructions win. Your global accuracy rule doesn’t apply inside a project unless you add it there.

Try it now: a 10-minute setup

Paste the accuracy block into your custom instructions. Open Memory and delete anything that’s no longer true. Then ask ChatGPT one question about something current in your work, once with Instant and once with a thinking level and search on, and compare the two answers. You’ll see which setting matters most for how you use it.

If the errors themselves still puzzle you, why AI makes things up explains what’s going on under the hood.

Related guides

• The AI Fact-Check Workflow: Verify Any Answer in 3 Minutes
• How to Improve AI Accuracy: 7 Fixes That Actually Work
• Your AI Is Lying to You: 3 Prompts to Catch Hallucinations
• Free AI Hallucination Checker: A Prompt That Flags Made-Up Facts
• The AI Memory Kit